V‑Blaze version 7.5.0-3 (October 2023)

V‑Blaze version 7.5.0-3, voci-webapi-3.1.1-1

New features and improvements

• [SPM-2526] Added new Out-of-Vocabulary (OOV) ASR flow.

  OOV (out-of-vocabulary) is an ASR tuning feature designed to improve transcription accuracy for audio that contains brand- and industry-specific terminology. OOV enhances existing language models with new words and preferential treatment for those words.

  • Only language models version 7.5 and higher support OOV.

  • [SPM-2581] Added new stream tag oov for user-provided OOV lists.

  • [SPM-2583, 3188] OOV lists are cached to optimize performance.

  • OOV has fully deprecated Hinting as a tuning feature.

• [SPM-3392->3700] Improved date detection and formatting, and added logic to protect dates from numerical redaction.

• [SPM-3359] subst_rules now supports splitting rules on a comma in addition to newline characters.

• [SPM-3818] senti_rules now supports splitting rules on a comma in addition to newline characters.

• [SPM-3889] Two-digit number sequences beginning with zero are now concatenated in transcript output. For example, two-digit sequences previously transcribed as 0 2, 0 3 are now output as 02, 03.

• [SPM-3826] Manual sentiment is now enabled for all language models. Previously, manual sentiment was restricted to English.

• [INTERNAL ONLY] [SPM-3864] Improved sentiment detection.

  • Expanded the range that limits utterance-phrase capacity for sentiment detection. Previously, only utterances containing 2,3,4, or 5 words were evaluated for sentiment. Now, utterances containing as few as 1 or as many as 10 words are evaluated for sentiment.

  • Improved consistency of cursor position after detecting a sentiment match.

  • Improved search loop logic.

  • Improved computation logic for sentimentex details.

  • [SPM-3890] The sentiment plugin now accepts Unicode characters.

• [SPM-3666] Improved memory footprint for decodeworker processes by resetting them after the maximum number of streams has been processed.

  [INTERNAL ONLY] The default maximum number of streams is 100 and is configurable with the workercycle system configuration setting.

Bug fixes

• [SPM-3669] Resolved an issue that caused number translation to fail on uppercase internal substitutions.

• [INTERNAL ONLY] [SPM-3786] Resolved an issue with frame calculation when transcribing dates in French. When transcribing French dates, calculated frames are now divided evenly across all utterance components.

• [SPM-3825] Resolved an issue that caused the senti_list tag to fail when processing stereo audio.

• [SPM-3934] Resolved an issue with sentiment calculation when sentiment-bearing phrases cross utterance boundaries. Sentiment for those words and phrases is now properly set.

• [SPM-4017] Resolved issues with Dutch models. Added the connecting word ën and resolved errors with hour and month conditional phrases.

[INTERNAL ONLY] releaseinfo.txt

20230718 v7.5.0-1
  - added new OOV flow and disabled hinting
    - new stream tags oov{,adjust,boost}
    - added ramfs oov and streams subdirs
    - added oov.{cache,streams,errors} to status
  - sentiment
    - support manual sentiment (senti_rules/list) for all languages
    - support manual sentiment phrases from 1 to 10 words (was 2-5)
    - removed nested list when empty sentimentex
    - fixed handling of unicode in sentiment computation
    - fixed occasional drop of sentimentex with merged utts
    - fixed senti_list handling of multiple channels 
  - textproc
    - support comma separator for subst/senti_rules
    - cap substitution backrefs at \9 so can be followed by digit
    - improve date detection and exclude dates from scrubbing
    - numtrans improvements
      - eng-gb pound/quid/pence, fre dashed number words, spa comma/decimals
  - webapi
    - added metrics for oov_{cache,streams,errors}
    - rename debug-conn to debugconn
  - updated level VAD defaults to DtT recommended settings
    - activitylevel/uttmaxsilence/uttminactivity/uttpadding
  - discard init logs unless debuglogs=true
  - updated word end time +1 frame to match sphinx
  - tweaked top-level emotion flow split point +1
  - licensing
    - added sequence number to usage request
    - fixed devel=true support
  : voci-server-{client,server{,-sw},service,sentiment,textproc}-7.5.0-1
  : voci-{server{,-core}-{gpu{9,10,11},sw},lmtk}-7.5.0-1
  : voci-server-{asrlicense,licensemgr-{gpu,sw}}-7.5.0-1
  : voci-webapi-3.1.1-1

[INTERNAL ONLY] releaseinfo 
20230802 v7.5.0-2
  - improved scrubconf noscrub processing to handle missing noscrub section
    - fixes dates like "May 14th @@####@@"
  - add oov and subst_rules to usage tracking
  - add simple OOV validation
    - open paren and marker/filler prefixes (</++) not allowed
    - other symbols or digits (except _-') require soundout
  - escape OOV cased subs that have special textproc suffixes
  - fix warning when no OOV models installed
  - service - if systemd always remove lock / force start
  - fix service/serverstatus usage query
  - textproc
    - loosen restrictions on non-letter chars in words
    - support /../ escaping subst search words with special chars
    - improve date handling and scrub protection
      - better support for connecting words (9th of May)
      - keep small ordinals of date (6th not sixth)
      - support day spoken as "oh <digit>"
  : voci-server-{client,server{,-sw},service,textproc}-7.5.0-2

 [INTERNAL ONLY] releaseinfo
 20231018 v7.5.0-3
  - add decodeworker cycling to control proc mem growth
    - configurable workercycle sys default of 100 (streams)
  - added timeout to cond.waits in case notify sig is lost(?)
    - might be cause of very rare init stream hangs
  - textproc
    - Dutch numtrans improvement - added connecting word ën
    - fixed month/hour conditionals
  : voci-server-{server{,-sw},textproc}-7.5.0-3